CN109993224B - GEO satellite shape and attitude identification method based on deep learning and multi-core learning - Google Patents

Abstract

The invention provides a GEO satellite shape and attitude identification method based on deep learning and multi-core learning, which comprises the following steps: acquiring OCS sequence data of a GEO satellite for one year; preprocessing OCS sequence data; constructing a C-RNN model for automatically extracting characteristics of OCS sequence data based on a deep learning network, wherein the deep learning network comprises a cyclic neural network and a convolutional neural network; training a C-RNN model to obtain a plurality of feature vectors of OCS sequence data; based on the multi-kernel learning technology, a plurality of kernel functions are used for mapping different features, and a support vector machine is used for identifying the shape and the posture of the satellite. Under the condition of no need of prior information, the method combines deep learning based on the cyclic neural network and the convolutional neural network and the multi-core learning technology based on the support vector machine, utilizes a pure data driving mode, and uses OCS sequence data to automatically identify the shape and the posture of the GEO satellite.

Description

Shape and Attitude Recognition Method of GEO Satellite Based on Deep Learning and Multi-core Learning

technical field

The invention belongs to the technical field of GEO satellite shape and attitude recognition, and in particular relates to a GEO satellite shape and attitude recognition method based on deep learning and multi-core learning.

Background technique

At present, the number of space objects is increasing, and space situational awareness (Space Situational Awareness) has become one of the important research topics in the world. Geosynchronous orbit (GEO) is an important space asset, and verifying the status of all satellites and objects in GEO is important to properly assess the GEO environment. At the same time, detailed space object characteristics (such as shape, attitude and motion state) can be used to accurately predict the trajectory and behavior of space objects, providing important information capability guarantee for space situational awareness. Optical observation is an important tool to obtain information on space objects. A large number of existing optical telescopes are still the main method for observing GEO targets in my country. More specifically, analysts can extract the shape of space objects through the photometric sequence data obtained by the optical observation system. , size, attitude, reflectivity and material characteristics, and finally determine the behavior and intention of the space target. At present, optical scattering cross section (OCS) is widely used to represent the optical scattering properties of space targets, and the shape and attitude of GEO targets can be effectively identified by the visible light scattering properties. Because the OCS is only affected by the space target geometry, surface material, attitude and the relative position of the sun-space target-observation station, but has nothing to do with the observation distance and observation system, and at the same time, the OCS sequence data and the photometric sequence data can be converted to each other, so , OCS sequence data is very suitable for feature recognition of spatial targets.

Currently, analysts have been able to manually identify the shape and attitude of satellites from photometric sequence data. In general, the most common identification methods are physical models (such as inversion methods based on the dihedral model) and filters (such as Kalman filtering). The point pairing method proposed based on the dihedral model requires the satellite body to have Lambertian reflection characteristics and a complex three-dimensional shape. At the same time, it requires the satellite sail board to have specular and Lambertian reflection characteristics, and it is close to a plane structure. This method can quickly invert the GEO target. reflectivity and area. The Kalman filter method takes the characteristic parameters of the space target to be identified as the unknown state parameters of the system for optimal estimation. The shape of the space target and the direction of the inertial axis of the target can be identified by using the lossless Kalman filter. This method has better identification performance and faster speed. Also faster.

Although traditional physical models and filtering methods are fast, traditional methods require a lot of prior information and are greatly affected by the quality of the model. In addition, the data obtained from a large number of observation devices is so large that manual identification by humans is no longer feasible.

The method based on the dihedral model firstly requires a lot of structure of the target model, secondly, the method requires the same contribution of the reflectivity and area of the target body to each observation point, and the body pose of the two observation points must be the same. The method requires to maximize the time interval between point pairs or the position interval of the observation station, so as to improve the recognition effect of the model. Therefore, this method does not have generality and practicability at present.

The Kalman filtering method needs to fuse the phase angle data and the photometric sequence data. Therefore, certain prior knowledge is required. At the same time, the uncertainty of the model parameters of the spatial target itself will cause systematic errors in the Kalman filtering estimation.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide a GEO satellite shape and attitude recognition method based on deep learning and multi-core learning, which combines the Recurrent Neural Network (RNN) and Convolutional Neural Network without prior information. Deep learning (Convolutional Neural Network, CNN) and multi-core learning technology based on Support Vector Machine (SVM), using pure data-driven mode, use OCS sequence data to automatically identify the shape and attitude of GEO satellites.

In order to achieve the above object, the present invention is specifically realized through the following technical solutions:

The present invention provides a GEO satellite shape and attitude recognition method based on deep learning and multi-core learning, including:

Step 1: Obtain the optical scattering cross section (OCS) sequence data of the GEO satellite;

Step 2: Preprocess the OCS sequence data;

Step 3, building a C-RNN model for automatic feature extraction of OCS sequence data based on a deep learning network, where the deep learning network includes a cyclic neural network and a convolutional neural network;

Step 4: Train the C-RNN model to obtain multiple feature vectors of the OCS sequence data;

Step 5: Based on the multi-kernel learning technology, use multiple kernel functions to map different features, and use the support vector machine to recognize the shape and attitude of the satellite.

The method for obtaining the OCS sequence data of the GEO satellite in the step 1 includes:

The OCS sequence data of the space target is obtained by one or more methods of numerical calculation, photometric sequence data obtained from actual observation, and/or laboratory simulation measurements.

Further, the method for obtaining the OCS sequence data of the spatial target by means of numerical calculation includes: using the BRDF model and the OCS calculation method to obtain the OCS sequence data of the spatial target;

The photometric sequence data obtained through actual observation can be converted into OCS sequence data. The conversion relationship between photometric sequence data and OCS sequence data is:

Among them, m is the photometric sequence data (magnitude), and r is the distance between the observation equipment and the space target.

In the second step, the method for preprocessing the OCS sequence data includes:

According to the observation geometric position relationship corresponding to the OCS sequence data obtained in different observation intervals, the OCS sequence data obtained under different observation geometric position relationships are divided into different subsets, and the label of each subset is set to the corresponding A subclass of the category; the observed geometric positional relationship is the positional relationship of the sun, the space target, and the station.

In the third step, the method for constructing a C-RNN model for automatic feature extraction of OCS sequence data includes:

The C-RNN model consists of an encoder, a decoder and a classifier;

The encoder is used to take the OCS sequence data as input and generate a fixed-length feature vector as output;

The decoder reconstructs the input OCS sequence data through the feature vector generated by the encoder;

The classifier consists of three fully connected layers using the ReLU activation function and one output layer using the sigmoid activation function; the feature vector generated by the encoder is used as input; the sigmoid function is used to map the feature vector to the category as output, and the GEO satellite is obtained. Input the shape and pose corresponding to the OCS sequence data.

Further, the loss function of the C-RNN model is:

L=MSE+loss

Among them, MSE is the loss function used for OCS sequence data reconstruction, and loss is the loss function used in the shape and pose classification process; the back-propagation and gradient descent methods are used in the C-RNN model training process to minimize the total loss.

Further, the encoder includes two 1-D convolutional layers with Rectified Linear Unit (ReLU) activation function, and the ReLU function can be expressed as:

Among them, x is the input value; 1-D convolution uses a one-dimensional convolution kernel to convolve the OCS sequence data, thereby extracting the feature vector of the OCS sequence data; 1-D convolution is defined as:

为输入的OCS序列数据；n为序列数据的长度；W和k分别表示1-D卷积核和滑动步数；

为卷积后的输出向量；m＝n-n_k+1，n_k为卷积核的大小；in,

is the input OCS sequence data; n is the length of the sequence data; W and k represent the 1-D convolution kernel and the number of sliding steps, respectively;

is the output vector after convolution; m=nn _k +1, n _k is the size of the convolution kernel;

Each convolutional layer is followed by a dropout layer; the second convolutional layer is followed by a flatten layer, which converts multi-dimensional features into 1-D features; the final feature vector with the specified length is obtained by flattening the The outputs are passed to two fully connected layers, which use the ReLU activation function.

Further, in the decoder, two Gated Recurrent Unit (GRU) networks are applied to complete the reconstruction task of the OCS input signal;

The decoder takes the difference Δt _N between the feature vector and the sampling time as input, where N is the number of sampling points; copies the feature vector l times, where l is the set length of the output sequence of the decoder; The difference is also replicated l times; the feature vector is used to characterize the OCS sequence data, and the sampling time is used to determine the position of each point in the reconstructed sequence.

In the step 4, the method for training the C-RNN model to obtain multiple feature vectors of the OCS sequence data includes:

Input the OCS sequence data of length 200 into the C-RNN model; the C-RNN model needs to be trained 2 times, the first time a convolution kernel of size 5 is applied to two CNN layers; the second time the size is 3 The convolution kernel of C-RNN is applied to two CNN layers, and the number of iterations and batch size are set to 2000 and 1000 respectively. After each training, the feature vector output by the C-RNN model is saved;

Among them, the two CNN layers in the model have 60 filters and 100 filters respectively; the output of the encoder is a feature vector with an embedding size of 64;

The decoder uses a 2-layer bidirectional GRU layer of unit size 100; the input to the decoder is a feature vector of length 65, which includes the output of the encoder (length 64) and the sample time difference between the OCS sequence data points ( length is 1);

The classifier processes the feature vector generated by the encoder and gives the classification result of each OCS sequence data;

The Adam optimizer is used for network optimization with a learning rate of 1×10 ^-3 ; the dropout rate of each dropout layer is set to 0.25.

In the step 5, the method for identifying the shape and attitude of the satellite includes:

Based on the MKL technology, the feature vector generated by the C-RNN model using different convolution kernels is used as the input data of the support vector machine (SVM), and the basic kernel functions are combined by the multi-kernel linear combination method, and the multi-kernel linear combination is the final kernel. The function is used as the kernel function of SVM, and SVM is used for classification. The multi-kernel linear combination can be described as follows:

为特征空间，

为第i个归一化的基本核函数，K(x,z)表示由n个基本核函数线性组合而成的最终核函数，β_i表示第i个系数；基本核函数为多项式核函数，多项式内核可表示为：Among them, x, z∈X,

is the feature space,

is the ith normalized basic kernel function, K(x, z) represents the final kernel function formed by linear combination of n basic kernel functions, β _i represents the ith coefficient; the basic kernel function is a polynomial kernel function, The polynomial kernel can be expressed as:

为特征空间；R和d分别为常数和多项式的阶；其中，MKL为多特征融合方法，SVM为分类模型。目前，常用的多核学习(MKL)是一种基于SVM的多特征融合方法。通常，SVM是单内核的，很难选择最合适的核函数以及相应的参数来得到最佳的分类效果。而MKL是针对不同的特征采用不同的内核，为不同的内核分配不同的权重，然后训练每个核的权重，并选择核函数的最佳组合来完成分类任务。where: x,z∈X,

is the feature space; R and d are the order of a constant and a polynomial, respectively; among them, MKL is a multi-feature fusion method, and SVM is a classification model. Currently, the commonly used multi-kernel learning (MKL) is a multi-feature fusion method based on SVM. Usually, SVM is a single kernel, it is difficult to choose the most suitable kernel function and corresponding parameters to get the best classification effect. MKL uses different kernels for different features, assigns different weights to different kernels, then trains the weights of each kernel, and selects the best combination of kernel functions to complete the classification task.

The beneficial effects of the present invention are:

The present invention uses neural network to perform automatic feature extraction on OCS sequence data of GEO satellite; the C-RNN model proposed by the present invention includes an encoder composed of CNN, a decoder composed of RNN, and a classifier composed of fully connected neural network. The main function of the classifier in the C-RNN architecture is to maximize the distance between the features, rather than classifying; the present invention is to recognize the shape and attitude of the GEO satellite through multi-core learning technology and SVM. The way of learning is to complete the linear combination of multiple polynomial kernels. The invention can automatically perform feature extraction on the obtained OCS sequence data, which saves a lot of labor costs; the extracted features contain richer information of the OCS sequence data, and the classification effect can be improved during classification; the use of multi-core learning can integrate different features, make the classification results more accurate.

Description of drawings

Figure 1 shows a schematic diagram of the structure of the C-RNN model.

Figure 2 shows a schematic diagram of the principle of MKL.

Figures 3a to 3e show schematic diagrams of five satellite models.

Figures 4a to 4e are schematic diagrams showing the reconstruction results of the OCS sequence data of five satellites in attitude 2.

Figure 5 is a schematic diagram of the results of classifying OCS sequence data using a C-RNN structure with a convolution kernel of size 5.

Figure 6 is a schematic diagram of the results of classifying OCS sequence data using a C-RNN structure with a convolution kernel of size 3.

FIG. 7 is a schematic diagram of the classification result of the test OCS sequence data based on the support vector machine MKL.

Figure 8 is a schematic diagram of the training loss and validation loss of ENDECLA-CR, ENDE-CR, and ENDE-RR.

Detailed ways

In order to make those skilled in the art better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only Embodiments are part of the present invention, but not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by persons of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

Example 1

The embodiments of the present invention provide a method for recognizing the shape and attitude of a GEO satellite based on deep learning and multi-core learning, including:

Step 1: Obtain the Optical Scattering Cross Section Characteristic (OCS) sequence data of the GEO satellite.

The method for acquiring the OCS sequence data of the GEO satellite includes: acquiring the OCS sequence data of the space target by one or more of numerical calculation, photometric sequence data obtained by actual observation and/or laboratory simulation measurements. The method of obtaining the OCS sequence data of the space target by means of numerical calculation, including: using the BRDF model and the OCS calculation method to obtain the OCS sequence data of the space target; the photometric sequence data obtained through actual observation can be converted into OCS sequence data, photometric sequence data The conversion relationship with OCS sequence data is:

Among them, m is the photometric sequence data (magnitude), and r is the distance between the observation equipment and the space target.

Step 2: Preprocess the OCS sequence data; including:

According to the observation geometric position relationship corresponding to the OCS sequence data obtained in different observation intervals, the OCS sequence data obtained under different observation geometric position relationships are divided into different subsets, and the label of each subset is set to the corresponding A subclass of the category; the observed geometric positional relationship is the positional relationship of the sun, the space target, and the station. For example, if the OCS sequence data of target T in category C corresponds to n major observational geometric positional relationships (n≤5), then the OCS sequence data of target T in category C will be divided into n subcategories, and each subcategory will be divided into n subcategories. The labels of the classes are set to C1, C2, ..., Cn. Therefore, one year's OCS sequence data will be divided into N _C *n categories, where N _C is the number of categories. The above processing method can improve the training effect in the fourth step. In addition, each OCS sequence was processed into sequence data of length 200 for identification and classification.

Step 3, building a C-RNN model for automatic feature extraction of OCS sequence data based on a deep learning network, where the deep learning network includes a cyclic neural network and a convolutional neural network;

In order to use OCS data to identify the shape and attitude of GEO satellites, it is necessary to compress the entire OCS sequence data into feature vectors. The C-RNN feature extraction model is shown in Figure 1. M and D in the figure represent the dimensions of the CNN convolutional layers. The size of each convolutional layer is determined by the given convolution kernel size, the number of filters and the size of the input data. The length is determined, and the dropout layer is omitted in the figure.

The C-RNN model consists of an encoder, a decoder and a classifier; the encoder is mainly composed of CNN, which takes the OCS sequence as input and generates a fixed-length feature vector as output. The decoder reconstructs the input OCS sequence data through the feature vector generated by the encoder; the classifier consists of three fully connected layers using the ReLU activation function and one output layer using the sigmoid activation function; using the feature vector generated by the encoder as input ; Use the sigmoid function to map feature vectors to categories as output to obtain the shape and pose corresponding to the GEO satellite input OCS sequence data.

Since the proposed model contains two outputs, two loss functions need to be defined. The loss function of the C-RNN model is:

L=MSE+loss

Among them, MSE is the loss function used for OCS sequence data reconstruction, and loss is the loss function used in the shape and pose classification process; the back-propagation and gradient descent methods are used in the C-RNN model training process to minimize the total loss. The loss function used for OCS sequence reconstruction is mean square error (MSE). MSE can be given by:

为第i个OCS序列，

为第i个重构序列，w_i为权重系数，n表示输出序列的长度，N表示输入OCS序列的总数。where:

is the ith OCS sequence,

is the ith reconstruction sequence, w _i is the weight coefficient, n represents the length of the output sequence, and N represents the total number of input OCS sequences.

Use binary cross-entropy as the loss function for the shape and pose classification process. It should be noted that when binary cross entropy is used as the loss function, the labels of sequence data need to be binarized, and the output of the classifier is also a vector composed of 0s and 1s. The binary cross-entropy is given by:

和

分别为二值化后，标签中的第i个数值和分类器输出的类别向量中的第i个值，n_c表示分类器输出向量的长度。where:

and

After binarization, the i-th value in the label and the i-th value in the category vector output by the classifier are respectively, n _c represents the length of the classifier output vector.

The encoder is mainly composed of a CNN, which takes the OCS sequence as input and generates a fixed-length feature vector as output. Including two 1-D convolutional layers with Rectified Linear Unit (ReLU) activation function, the ReLU function can be expressed as:

Among them, x is the input value; 1-D convolution uses a one-dimensional convolution kernel to convolve the OCS sequence data, thereby extracting the feature vector of the OCS sequence data; 1-D convolution is defined as:

为输入的OCS序列数据；n为序列数据的长度；W和k分别表示1-D卷积核和滑动步数；

为卷积后的输出向量；m＝n-n_k+1，n_k为卷积核的大小；为了防止神经网络过度拟合，每个卷积层之后都有一个dropout层；第二个卷积层后有一层flatten层，该层将多维特征转换为1-D特征；最终得到的具有指定长度的特征向量是通过将flatten层的输出传递到两个全连接层产生的，全连接层使用ReLU激活函数。in,

is the input OCS sequence data; n is the length of the sequence data; W and k represent the 1-D convolution kernel and the number of sliding steps, respectively;

is the output vector after convolution; m=nn _k +1, n _k is the size of the convolution kernel; in order to prevent the neural network from overfitting, there is a dropout layer after each convolution layer; the second convolution layer This is followed by a flatten layer, which converts multi-dimensional features into 1-D features; the resulting feature vector with the specified length is generated by passing the output of the flatten layer to two fully connected layers, which are activated using ReLU function.

In the decoder, two Gated Recurrent Unit (GRU) networks are applied to complete the reconstruction task of the OCS input signal; GRU is a variant of RNN, which can effectively solve the long-term dependency problem of RNN, and in processing Outperforms standard RNNs on tasks with time series data.

The decoder uses the feature vector and the difference between the sampling time Δt _N as input, where N is the number of sampling points; copies the feature vector l times, where l is the set output sequence length of the decoder; The difference is also replicated l times; the feature vector is used to characterize the OCS sequence data, and the sampling time is used to determine the position of each point in the reconstructed sequence. By appending the time difference between sampling points as input, the C-RNN model can be used to process non-uniformly sampled time series data.

The classifier in the C-RNN model can be used to deal with multi-label classification problems, and the shape and pose of GEO objects can be directly classified using the classifier. However, in the feature extraction C-RNN model, the classifier is mainly used to maximize the distance between feature vectors corresponding to different targets and different poses, so that the feature information extracted by the encoder is more abundant and accurate. The classifier consists of three fully connected layers using ReLU activation function and one output layer using sigmoid activation function. The feature vector produced by the encoder is the input to the classifier. The output layer uses the sigmoid function to map features to categories. Through the classifier, the shape and pose corresponding to the input OCS sequence data of the GEO target can be directly obtained.

Step 4: Train the C-RNN model to obtain multiple feature vectors of the OCS sequence data; including:

The input layer of the C-RNN model will process OCS sequence data of length 200. Input the preprocessed OCS sequence data of length 200 into the C-RNN model; the C-RNN model needs to be trained 2 times, the first time a convolution kernel of size 5 is applied to two CNN layers; the second time Apply a convolution kernel of size 3 to two CNN layers, set the number of iterations and batch size to 2000 and 1000, respectively, after each training, save the feature vector output by the C-RNN model;

Among them, the two CNN layers in the model have 60 filters and 100 filters respectively; the output of the encoder is a feature vector with an embedding size of 64;

The decoder uses a 2-layer bidirectional GRU layer of unit size 100; the input to the decoder is a feature vector of length 65, which includes the output of the encoder (length 64) and the sample time difference between the OCS sequence data points ( length is 1);

Since the classifier needs to deal with the multi-label classification problem, the label of each OCS sequence data needs to be binarized. The classifier will process the feature vector generated by the encoder and give the classification result of each OCS sequence data;

The Adam optimizer is used for network optimization with a learning rate of 1×10 ^-3 ; the dropout rate of each dropout layer is set to 0.25.

Step 5: Based on the multi-kernel learning technology, use multiple kernel functions to map different features, and use the support vector machine to recognize the shape and attitude of the satellite. include:

Based on the MKL technology, the feature vector generated by the C-RNN model using different convolution kernels is used as the input data of the support vector machine SVM, and the basic kernel functions are combined by the multi-kernel linear combination method. The multi-kernel linear combination can be described as follows:

为特征空间，

为第i个归一化的基本核函数，K(x,z)表示由n个基本核函数线性组合而成的最终核函数，β_i表示第i个系数；基本核函数为多项式核函数，多项式内核可表示为：Among them, x, z∈X,

is the feature space,

is the ith normalized basic kernel function, K(x, z) represents the final kernel function formed by linear combination of n basic kernel functions, β _i represents the ith coefficient; the basic kernel function is a polynomial kernel function, The polynomial kernel can be expressed as:

为特征空间；R和d分别为常数和多项式的阶；其中，MKL为多特征融合方法，SVM为分类模型。将最终核函数用于SVM分类器，从而确定GEO目标的形状和姿态。多核学习的原理示意图如图2所示。where: x,z∈X,

is the feature space; R and d are the order of a constant and a polynomial, respectively; among them, MKL is a multi-feature fusion method, and SVM is a classification model. The final kernel function is used in the SVM classifier to determine the shape and pose of the GEO object. A schematic diagram of the principle of multi-core learning is shown in Figure 2.

The beneficial effects of the present invention are:

The present invention uses neural network to perform automatic feature extraction on OCS sequence data of GEO satellite; the C-RNN model proposed by the present invention includes an encoder composed of CNN, a decoder composed of RNN, and a classifier composed of fully connected neural network. The main function of the classifier in the C-RNN architecture is to maximize the distance between the features, rather than classifying; the present invention is to recognize the shape and attitude of the GEO satellite through multi-core learning technology and SVM. The way of learning is to complete the linear combination of multiple polynomial kernels. The invention can automatically perform feature extraction on the obtained OCS sequence data, which saves a lot of labor costs; the extracted features contain richer information of the OCS sequence data, and the classification effect can be improved during classification; the use of multi-core learning can integrate different features, make the classification results more accurate.

In order to verify the effect of the present invention, the Chinese Lijiang Observatory (25.6°N, 101.1°E, 2.465Km) was used to calculate the one-year OCS sequence data of 5 different satellites under 3 different attitudes (three-axis stabilization mode), The model structures of the five spatial targets are shown in Fig. 3a-Fig. 3e.

The acquired OCS sequence data can be divided into 15 categories, namely: (1) target 1, attitude 1; (2) target 1, attitude 2; (3) target 1, attitude 3; (4) target 2, attitude 1; (5) target 2, attitude 2 ; ⑹ target 2, posture 3; ⑺ target 3, posture 1; ⑻ target 3, posture 2; ⑼ target 3, posture 3; ⑽ target 4, posture 1; ⑾ target 4, posture 2; ⑿ target 4, posture 3; ⒀ Goal 5, Attitude 1; ⒁ Goal 5, Attitude 2; ⒂ Goal 5, Attitude 3. The three attitudes are: 1) the x-axis points to the satellite velocity direction, the z-axis faces the ground, the x, y, and z-axes are orthogonal to each other and satisfy the right-hand rule; 2) the y-axis points to the satellite velocity direction, and the z-axis is along the direction of the solar panel , x, y, z axes are orthogonal to each other and satisfy the right-hand rule; 3) the z-axis points to the direction of the satellite velocity, the x-axis points to the earth's center, and the x, y, z axes are orthogonal to each other and satisfy the right-hand rule.

According to step 1, 6,915 OCS sequence data (15 types in total) of five GEO satellites in three attitudes can be obtained, 1,383 for each satellite, and each photometric sequence data corresponds to a different observation interval. OCS sequence data less than 200 in length were removed, so the total available data was 5,505, 1,101 for each satellite, and 367 for each attitude. According to the observation geometric position relationship in the observation interval, the OCS sequence data corresponding to each satellite is divided into 5 sub-categories (that is, 5 main observation geometric position relationships): the first sub-category has 129 OCS sequence data; the second sub-category has 420 pieces of OCS sequence data; the third subclass has a total of 141 OCS sequence data; the fourth subclass has a total of 399 OCS sequence data; the fifth subclass has a total of 12 OCS sequence data. To make the numerically computed OCS data more reliable, random error values that satisfy the Gaussian distribution are added to the original OCS data. About 70% of the data of the OCS sequence dataset is extracted as the training set, and the rest of the data are used as the validation set and test set, respectively. After data preprocessing, the number of OCS sequences in the training set is about 12,500.

The training data is input into the built model, and after 2000 training iterations, the trained model is obtained. Taking about 2,160 pieces of OCS sequence data as test data, the test data is divided into 15 categories with 144 pieces of data in each category (each category represents one satellite in one attitude). Use the trained model to reconstruct and classify the test data. In the first training, the decoder reconstructs the original test OCS sequence data at 5 satellites and attitude 2, and the results are shown in Figure 4a-Figure 4e . In the figure, "Sat1" indicates "Satellite 1", and "Attitude TWO" indicates the second attitude among the three satellite attitudes.

After two trainings, the classification results of the C-RNN architecture classifier are shown in Figure 5 and Figure 6. In the figure, "T0A1" represents satellite 1, attitude 1, and the value on the diagonal is the number of correct classifications. When the encoder uses a convolution kernel of size 5, the classifier achieves a classification accuracy of 91.9%; when using a convolution kernel of size 3, the classifier achieves a classification accuracy of 83.3%. Due to the different sizes of the convolution kernels, the encoder will generate different feature vectors, so the classification results are also different. With the model structure unchanged, using an encoder with a convolution kernel of size 5, the features obtained from the encoder can better represent the OCS sequence data of the three attitudes of five satellites.

To use multi-kernel learning for classification, set R to 0 in the basic polynomial kernel function. According to the obtained two kinds of features, several polynomial kernels are constructed in turn, wherein the order of each polynomial kernel function is d=1, 2, 3, . . . , 10, respectively. Taking the support vector machine (SVM) as the classifier, the final combined kernel K is used as the kernel function of the support vector machine, and the training set is processed for training. After the classifier is trained, the optimal penalty value C of the SVM is 1000, and the linear combination coefficients of the kernel function are all 0. The results of MKL classification based on SVM are shown in Fig. 7. The classification accuracy obtained by multi-kernel learning reaches 99.58%.

To evaluate the feature selection performance of the proposed C-RNN architecture (referred to as ENDECLA-CR), several network models are first constructed to compare with the proposed C-RNN architecture. The first network model is a model that removes the classifier in C-RNN (denoted this model as ENDE-CR); the second network model is to replace the two convolutional layers of the encoder in ENDE-CR with a size of 2 GRU layers of 100 (this model is called ENDE-RR). The model parameters and training parameters used by these two models are consistent with those used by the proposed C-RNN model. The training data used to train the C-RNN model is used as the input training data for the ENDE-CR and ENDE-RR models. After training, the training loss and validation loss of ENDECLA-CR, ENDE-CR, and ENDE-RR are shown in Figure 8.

At the same time, Table 1 lists the classification accuracy rates of the present invention and the other six feature extraction models for comparison, showing the superior recognition performance of the present invention. The remaining 6 feature extraction models include Principal Component Analysis (PCA), Linear Discriminant Analysis (LDA), Dictionary Learning (DL) and ENDE-CR, ENDE-RR and a simple deep neural network (ENDE-RR for encoder and decoder is replaced by a simple multi-layer fully connected structure, this model is called SDNN). The SDNN model contains 1 input layer (200 units), 4 fully connected layers (500 units per hidden layer) and 1 output layer (200 units). The number of training iterations of SDNN is set to 2000, the loss function uses MSE, and the optimizer uses Adam.

The training and test sets used in the comparison model are the same as those used in the C-RNN model. The features extracted by the trained 7 models are separately fed into a support vector machine with linear kernels for classification. The reason for using this classifier is that it does not map the input features to a high-dimensional space or perform other transformations, which can guarantee the validity of the comparison results.

The recognition performance of the model is evaluated by the average precision MAP of each model classification. MAP is the average prediction precision of all categories and can be expressed as:

where N is the total number of categories, precision _c is the classification accuracy of category c, x _Tc is the number of correctly predicted OCS sequences of category c, and Mc is the total number of OCS sequences with the recognition result of category c. The k in the table represents the size of the convolution kernel used by the C-RNN structure.

Table 1

According to the classification performance of the 7 models in Table 1, the recognition performance of PCA and LDA models reach 61% and 76%, respectively; compared with PCA and LDA, DL has better recognition performance, and the recognition accuracy rate is 73%; The recognition performance of SDNN is better than that of traditional feature extraction methods such as PCA, LDA, and DL, and the recognition accuracy rate reaches 84.5%; ENDE-CR and ENDE-RR have great advantages in feature extraction, and the recognition accuracy rate reaches 95.8% and 83%, respectively. %. From the results, it can be seen that the recognition accuracy of the proposed C-RNN structure is the best. When the kernel size is 3 and 5, the recognition accuracy exceeds 98%. The classifier in C-RNN increases the distance between features of different categories, so that the feature vector generated by the encoder can better represent the OCS sequence data, making the proposed C-RNN architecture have good classification performance.

The above-mentioned serial numbers of the embodiments of the present invention are only for description, and do not represent the advantages or disadvantages of the embodiments.

It should be noted that, for the sake of simple description, the foregoing method embodiments are all expressed as a series of action combinations, but those skilled in the art should know that the present invention is not limited by the described action sequence. As in accordance with the present invention, certain steps may be performed in other orders or simultaneously. Secondly, those skilled in the art should also know that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required by the present invention.

The above are only specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed by the present invention. should be included within the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims (9)

1. a GEO satellite shape and attitude recognition method based on deep learning and multi-core learning, is characterized in that, comprises: Step 1: Obtain the optical scattering cross section (OCS) sequence data of the GEO satellite; Step 2: Preprocess the OCS sequence data; Step 3, building a C-RNN model for automatic feature extraction of OCS sequence data based on a deep learning network, where the deep learning network includes a cyclic neural network and a convolutional neural network; Step 4: Train the C-RNN model to obtain multiple feature vectors of the OCS sequence data; Step 5. Based on the multi-kernel learning technology, use multiple kernel functions to map different feature vectors, and use the support vector machine to identify the shape and attitude of the satellite; In the third step, the method for constructing a C-RNN model for automatic feature extraction of OCS sequence data includes: the C-RNN model is composed of an encoder, a decoder and a classifier; the encoder is used to extract the OCS sequence data. As input, a fixed-length feature vector is generated as output; the decoder reconstructs the input OCS sequence data through the feature vector generated by the encoder; the classifier consists of three fully connected layers using the ReLU activation function and one output using the sigmoid activation function Layer composition; use the feature vector produced by the encoder as input; use the sigmoid function to map the feature vector to the category as output, obtain the distance between maximized features. 2. The method of claim 1, wherein the method for obtaining the OCS sequence data of the GEO satellite in the step 1 comprises: The OCS sequence data of the space target is obtained by one or more methods of numerical calculation, photometric sequence data obtained from actual observation, and/or laboratory simulation measurements. 3. method as claimed in claim 2 is characterized in that, the method that obtains the OCS sequence data of space target by the mode of numerical calculation, comprises: use BRDF model and OCS calculation method to obtain the OCS sequence data of space target; The photometric sequence data obtained through actual observation can be converted into OCS sequence data. The conversion relationship between photometric sequence data and OCS sequence data is:

Among them, m is the photometric sequence data, and r is the distance between the observation equipment and the space target. 4. The method of claim 1, wherein in the step 2, the method for preprocessing the OCS sequence data comprises: According to the observation geometric position relationship corresponding to the OCS sequence data obtained in different observation intervals, the OCS sequence data obtained under different observation geometric position relationships are divided into different subsets, and the label of each subset is set to the corresponding A subclass of the category; the observed geometric positional relationship is the positional relationship of the sun, the space target, and the station. 5. The method of claim 1, wherein the loss function of the C-RNN model is: L=MSE+loss Among them, MSE is the loss function used for OCS sequence data reconstruction, and loss is the loss function used in the shape and pose classification process; the back-propagation and gradient descent methods are used in the C-RNN model training process to minimize the total loss. 6. The method of claim 1, wherein the encoder includes two 1-D convolutional layers with Rectified Linear Unit (ReLU) activation functions, and the ReLU function can be expressed as:

Among them, x is the input value; 1-D convolution uses a one-dimensional convolution kernel to convolve the OCS sequence data, thereby extracting the feature vector of the OCS sequence data; 1-D convolution is defined as:

in,

is the input OCS sequence data; n is the length of the sequence data; W and k represent the 1-D convolution kernel and the number of sliding steps, respectively;

is the output vector after convolution; P=nn _k +1, n _k is the size of the convolution kernel; Each convolutional layer is followed by a dropout layer; the second convolutional layer is followed by a flatten layer, which converts multi-dimensional features into 1-D features; the final feature vector with the specified length is obtained by flattening the The outputs are passed to two fully connected layers, which use the ReLU activation function. 7. The method of claim 1, wherein, in the decoder, two gated recurrent unit (GRU) networks are applied to complete the reconstruction task of the OCS input signal; The decoder takes the difference Δt _N between the feature vector and the sampling time as input, where N is the number of sampling points; copies the feature vector l times, where l is the set length of the output sequence of the decoder; The difference of is also replicated l times; the feature vector is used to characterize the OCS sequence data, and the sampling time is used to determine the position of each point in the reconstructed sequence. 8. The method according to one of claims 1-7, wherein in the step 4, training a C-RNN model, and the method for obtaining a plurality of feature vectors of OCS sequence data comprises: Input the OCS sequence data of length 200 into the C-RNN model; the C-RNN model needs to be trained 2 times, the first time a convolution kernel of size 5 is applied to two CNN layers; the second time the size is 3 The convolution kernel of C-RNN is applied to two CNN layers, and the number of iterations and batch size are set to 2000 and 1000 respectively. After each training, the feature vector output by the C-RNN model is saved; Among them, the two CNN layers in the model have 60 filters and 100 filters respectively; the output of the encoder is a feature vector with an embedding size of 64; The decoder uses a 2-layer bidirectional GRU layer of unit size 100; the input to the decoder is a feature vector of length 65, which includes the output of the encoder of length 64, and the samples between the OCS sequence data points of length 1 time difference; The classifier processes the feature vector generated by the encoder and gives the classification result of each OCS sequence data; The Adam optimizer is used for network optimization with a learning rate of 1×10 ^-3 ; the dropout rate of each dropout layer is set to 0.25. 9. The method of claim 1, wherein in the step 5, the method for identifying the shape and attitude of the satellite comprises: Based on the multi-kernel learning (MKL) technology, the feature vector generated by the C-RNN model using different convolution kernels is used as the input data of the support vector machine (SVM), and the basic kernel functions are combined by the multi-kernel linear combination method. The multi-kernel linear combination can describe as follows:

Among them, x, z∈X,

is the feature space,

is the ith normalized basic kernel function, K(x, z) represents the final kernel function formed by linear combination of n basic kernel functions, β _i represents the ith coefficient; the basic kernel function is a polynomial kernel function, The polynomial kernel can be expressed as:

where: x,z∈X,

is the feature space; R and d are the order of a constant and a polynomial, respectively; among them, MKL is a multi-feature fusion method, and SVM is a classification model.

Images